Across 1.6 million visits, heavy AI-scribe users saved 2.5 note minutes per encounter. After-hours work did not change

One of the most practical claims made for generative AI is that it can give clinicians time back. Documentation is repetitive, compulsory and often carried home after clinic hours. A large real-world study published in Annals of Internal Medicine examined what happened when a Boston health system rolled out the Heidi AI scribe with few limits on access. The result is encouraging on one narrow, measurable outcome: clinicians who used it heavily spent less active time writing notes. It is much less decisive on the more ambitious hope that AI removes after-hours paperwork or increases patient-facing capacity.[1][2]

Researchers analyzed 1,637,227 patient encounters with notes between July 2024 and January 2026 in a 14-hospital system. There were 902 enrolled outpatient clinicians: 605 users and 297 who enrolled but did not use the scribe. An interrupted time-series design examined changes around each clinician's enrollment, accounting for measured patient characteristics and visit complexity. This was not a random assignment of clinicians to use or avoid AI. Differences in motivation, case mix and workflow could affect the observed relationship.[1][2]

The saved minutes were real within the study's measurement

Enrollment overall was associated with an immediate 1.21-minute reduction in active note-writing time per encounter, with a 95% confidence interval from 0.68 to 1.74 minutes saved. The association was larger for high-intensity users, at 2.50 minutes per encounter, with an interval of 1.45 to 3.55 minutes. Relative to enrolled nonusers, heavy users had a 2.20-minute greater reduction in the study's adjusted comparison. That pattern is consistent with benefit arising from active tool use rather than simply access to a subscription.[1][2]

The distinction between active note time and calendar time is essential. A reduction of 2.5 active minutes can matter across a busy clinic, but it does not automatically translate into a shorter shift, an additional appointment or less burnout. Those outcomes depend on how freed intervals are used, whether staff need to review errors, how many notes are drafted by the tool and whether the surrounding electronic-health-record workflow changes. A health system may absorb saved minutes into other tasks without altering its published hours or throughput.[1][2]

In this analysis, the researchers detected no difference in same-day note closure, after-hours closure or work relative value units, a billing-related productivity measure. Heavy users also produced longer notes: compared with nonusers, their notes averaged approximately 171 additional characters in the reported comparison. Length is not the same as quality. Extra details can improve documentation when they are correct and relevant, yet can also require more review or introduce distracting text.[1][2]

A large denominator does not create randomization

The study's 1.6 million encounters are a strength for observing real usage across many clinicians and time periods. They do not fix the causal limitations of an observational design. A clinician who chooses to use a scribe frequently may also be more willing to redesign note templates, delegate preparation or learn shortcuts. The authors themselves note that their data cannot distinguish which individual post-enrollment notes were AI-authored rather than clinician-authored. That affects how precisely any per-note improvement can be attributed to the technology. One author reported an affiliation with the scribe vendor Heidi Health; the documented affiliation is relevant when evaluating potential conflicts, without by itself invalidating the published analysis.[1][2]

A prior randomized trial involving 238 outpatient physicians and two different scribe products provides useful methodological context, although it tested other products and an earlier period. Its design illustrates why controlled allocation can help separate a technology effect from who elects to use it. The studies should not be pooled as though they are evaluations of one identical system. Vendor, interface, training, specialty and baseline documentation burden all influence the conditions of benefit.[1][2]

For procurement, the operational questions are concrete. How much active note time falls for clinicians who actually use the product? What proportion of generated notes need correction? Do patients consent appropriately to the recording workflow? Does after-hours work decline after allowing for changes in caseload? Who reviews the final note before it enters the legal medical record? A sponsor claiming thousands of hours saved should expose its denominator, the baseline and the counterfactual.[1][2]

The measured reduction is a legitimate human benefit: time that clinicians no longer spend typing during documented encounters. The larger promise—more attentive care, shorter waits or less exhausted clinicians—still requires the corresponding evidence. A useful AI-health story does not need to promise a transformed hospital when its strongest result is two and a half minutes returned to a professional's working day.[1][2]

Report an error or suggest a correction